SCARF: A Single Cell ATAC-seq and RNA-seq Foundation Model

SCARF is a large-scale foundation model designed for single-cell ATAC-seq and RNA-seq.
It provides pretrained weights, preprocessing pipelines, and tutorials to accelerate downstream biological discovery.

---

🚀 System Requirements

• Operating system: Linux (Ubuntu 20.04+)
• Python version: == 3.12.3
• Dependencies:
  • PyTorch >= 2.3.1
  • Scanpy >= 1.11.0
  • Anndata >= 0.9
  • scikit-learn == 1.5.2
  • transformers==4.46.3
  • numpy, pandas, matplotlib, seaborn, jupyter
• Hardware:
  • CPU: x86_64 architecture (tested on Intel i9 and AMD EPYC)
  • GPU (recommended): NVIDIA GPU with CUDA >= 11.8 (tested on A800, H100)
  • Minimum RAM: 40 GB

---

⚙️ Installation Guide

1. Clone the repository

git clone https://github.com/JiekaiLab/scarf.git
cd scarf

2. Create conda environment and install dependencies

CONDA_CHANNEL_PRIORITY=flexible conda env create -n scarf -f environment.yml

📊 Quick start

We provide example datasets and pretrained models for quick testing.

Download demo data and pretrained model files

Run the notebook (download_data.ipynb) to download automatically:

• Download the demo dataset (demo_hPBMC.tar.gz) into the data/ folder.

• Download model files (model_files.tar.gz) and extract:

  • weights/ → into the weights/ folder

  • prior_data/ → into the prior_data/ folder

This ensures all required data and weights are available locally.

Run SCARF on your own data

1. Preprocess your single-cell data (preprocess.ipynb)

• 600GB Memory required for preprocessing the sample data provided
• Expected runtime : ~6 hours

2. Run inference (embedding.ipynb)

• 10GB Memory required for inference the sample data provided.
• Expected runtime on a normal desktop (40GB RAM, no GPU): ~2–3 minutes
• Expected runtime on 1 GPU : ~20 seconds

🎯Downstream Tasks

We provide ready-to-use Jupyter notebooks demonstrating how to apply SCARF to different downstream tasks:

• Cell type prediction (CellType_prediction.ipynb)
  Predicts cell type labels from multi-omic embeddings.

• Cell Matching (Cell-matching.ipynb)
  Aligns and matches cells across modalities (scRNA-seq and scATAC-seq).

• Cell RNA-Inference (RNA-Inference.ipynb)
  Predicts gene expression of cells through scATAC-seq data. For this task, we provide precomputed embeddings (both RNA and ATAC) stored on :

  • RNA embeddings and reference anndata (preprocessed for inference,will be publicly released soon.)
  • ATAC embeddings (paired with RNA data,will be publicly released soon.)

📂 Repository Structure

SCARF/
├── data/                 # data for demo
├── downstream_tasks/     # Jupyter notebooks for demo and usage
├── scarf/                # model file
├── prior_data/           # Token dictionaries and metadata
├── scripts/              # Preprocessing and inference scripts
├── weights/              # Pretrained model weights (download from Zenodo)
└── environment.yml       # Dependencies

---

📜 License

This project is released under the GNU General Public License v3.0.
See LICENSE for details.

---

🔗 Links

• GitHub Repository: JiekaiLab/scarf
• Pretrained weights & large files:

---

📖 Citation

If you use SCARF in your research, please cite:

@misc{SCARF2025,
  title   = {SCARF: A Single Cell ATAC-seq and RNA-seq Foundation Model},
  author  = {Guole Liu#,Tianyu Wang#,Yingying Zhao#,Quanyou Cai#,Xiaotao Wang#,Ziyi Wen,Yaofeng Wang,Lihui Lin*, Yongbing Zhao*, Ge Yang*,Jiekai Chen*},
  year    = {2025},
  url     = {https://github.com/JiekaiLab/scarf},
  doi     = {https://doi.org/10.1101/2025.04.07.647689}
}

---